基于融合残差注意力机制的卷积神经网络地震信号去噪  

Seismic Signal Denoising Based on Convolutional Neural Network with Residual and Attention Mechanism

作　　者：刘霞[1] 孙英杰 Liu Xia;Sun Yingjie(School of Physics and Electronic Engineering,Northeast Petroleum University,Daqing 163318,Heilongjiang,China;School of Electrical Engineering&Information,Northeast Petroleum University,Daqing 163318,Heilongjiang,China)

机构地区：[1]东北石油大学物理与电子工程学院,黑龙江大庆163318 [2]东北石油大学电气信息工程学院,黑龙江大庆163318

出　　处：《吉林大学学报（地球科学版）》2023年第2期609-621,共13页Journal of Jilin University：Earth Science Edition

基　　金：黑龙江省自然科学基金项目(F201404)。

摘　　要：由于U型卷积神经网络(Unet)在地震数据去噪中存在计算量大、网络退化和泛化能力弱等问题,本文为了提高去噪效果以及增强模型的泛化性,提出了一种融合残差注意力机制的卷积神经网络(RAUnet)。该网络结构主要由编码和解码两部分构成,网络的每个卷积层之后都加入了批标准化和带泄露整流激活函数。在编码器中,为了提高对噪声的提取能力,引入了残差结构和卷积块注意力模块。残差结构利用残差跳跃连接的方式减弱了网络退化,降低了特征映射的难度。卷积块注意力模块使用通道和空间的混合注意力权重,能提升相关度高的特征并抑制相关度低的特征。在解码器中,为了提升特征融合的维度恢复能力,选用双线性插值方式进行上采样。实验测试结果表明,对于合成地震信号,本文方法对简单模型和复杂模型随机噪声的压制效果均更有效,并且更好地保护了有效信号;对于实际地震信号,本文方法仍然能在去噪的同时尽量保持有效信号中的细节,对叠前数据和叠后数据都展现出了良好的泛化性。Since U-shaped convolutional neural network(Unet)has problems such as large amount of computation,network degradation and weak generalization ability in seismic data denoising,this paper proposes a convolutional neural network(RAUnet)with residual and attention mechanism to improve the denoising effect and enhance the generalization of the model.The network structure is mainly composed of encoding and decoding,and batch normalization and Leaky ReLU activation functions are added to each convolutional layer of the network.In the encoder,in order to improve the ability to extract noise,a residual structure and a convolutional block attention module are introduced.The residual structure uses the residual skip connection to weaken the network degradation and reduce the difficulty of feature mapping.The convolutional block attention module uses a mix of channel and spatial attention weights,which boosts highly relevant features and suppresses less relevant ones.In the decoder,in order to improve the dimensional recovery ability of feature fusion,bilinear interpolation is used for up-sampling.Experimental results show that for synthetic seismic signals,the proposed method is more effective in suppressing random noise in both simple and complex models,and better protects the effective signals.For the filed seismic signal,the proposed method can still keep the details of the effective signal as much as possible while denoising,and shows good generalization for both pre-stack and post-stack data.

关 键 词：地震信号 深度学习 卷积神经网络(CNN) 去噪 

分 类 号：P631.4[天文地球—地质矿产勘探]